Extreme value estimation for correlated observations

Statistical modeling of extreme events is the object of interest in many application areas. When estimating such rare events from a time series, extreme value theory is commonly used. In that case, series with independent members are required. However, the assumption of independence is not satisfied in many situations. There are two approaches (block maxima, peaks-over-threshold) which result in series with independent members, but the length of the series is substantially reduced. In this paper, stationary series with short-time dependence described by the extremal index theta is considered, and two estimators of theta are introduced. Behavior of the estimators is assessed using simulations. The described methods are used in an analysis of real hydrological data, and compared with classical peaks-over-threshold approach.